search

ethz-asl / analog_gauge_reader

Open framework for reading analog gauges, deployable on real world robotic systems
MIT License
23 stars 4 forks source link

Feature/keypoint training #8

Closed mauritsreitsma closed 1 year ago

mauritsreitsma commented 1 year ago

Now do training with random crops. Setup that it is easy to extend with other augmentations. Now use Dinov2 as a Backbone.

kekeblom commented 1 year ago

How did this work? Any results or tests?

kekeblom commented 1 year ago

How did you validate that the feature extractor and data augmentation implementation works? Did you have some notebook or debugging code that you used to make sure that the resulting output is correct?

mauritsreitsma commented 1 year ago

How did you validate that the feature extractor and data augmentation implementation works? Did you have some notebook or debugging code that you used to make sure that the resulting output is correct?

Yeah i just had some debugging code that showed me the random crops, which looked good.

For the feature extractor i didn't have any code, tbh i didn't really know how to check if those are correct, since they are quite abstract. So there i just kind of hoped for the best.. do you have an idea to check if this is done correctly?

mauritsreitsma commented 1 year ago

How did this work? Any results or tests?

Yes this actually works quite nicely, i can send you some results via google chat

kekeblom commented 1 year ago

Did the changes make a difference?

mauritsreitsma commented 1 year ago

Did the changes make a difference?

Yes training results looks quite good now. i doubled resolution and swapped order of regression and upsampling and let the model run a bit longer, still not yet till convergence. I'll send results in the google chat. I have not tested it yet in the pipeline with on test images